Method of visualizing results of performance monitoring and analysis in a parallel computing system

Abstract

A method of visualizing the results of performance monitoring and analysis for a parallel computing system in which a plurality of processors execute a parallel processing program composed of a plurality of routines. First, information on the execution time of each routine is collected in a realtime manner while the program is concurrently executed by the plurality of processors. Second, a maximum, average, and minimum values of the execution time of each routine are calculated, based on the information collected for the plurality of processors. Third, the collected information is summarized as an execution profile and displayed in graphical form by using bar graphs, pie graphs, or radar charts. For each procedure or program loop, the present method clarifies the percentages of net user program execution, communication, synchronization, and other parallelization overheads, as well as indicating their deviations. This feature helps the user to understand the actual program behavior and to tune up the parallel processing program.

Claims

1. A method of visualizing results of performance monitoring and analysis for a parallel computing system in which a plurality of processors concurrently execute a parallel processing program composed of a plurality of routines, the method comprising:

collecting from the plurality of processors, execution time records for each routine being executed concurrently by the plurality of processors, in consideration of classification of the routines;

obtaining a maximum value, an average value, a minimum value, and standard deviation of the execution time records of each routine collected from the plurality of processors, based on the information collected for the plurality of processors; and

graphically displaying execution profile information including the maximum value, the average value, the minimum value, and the standard deviation of the execution time of each routine.

2. A method according to claim 1, wherein said collecting accumulates the execution time of each routine each time the routine is invoked and executed.

3. A method according to claim 1, wherein said collecting collects the execution time of each routine that is designated as a subject of profiling.

4. A method according to claim 1, wherein said collecting classifies the routines into user program, communication tasks, synchronization tasks, and other tasks.

5. A method according to claim 3, wherein the subject of profiling includes procedures, loops, and parallel processing library routines.

6. A method according to claim 1, wherein said graphically displaying extracts at least one of caller routines and callee routines from among the routines in the parallel processing program and displays the execution profile information pertaining to the routines extracted.

7. A method according to claim 1, wherein said graphically displaying uses a bar graph to graphically display the execution profile information of each routine.

8. A method according to claim 1, wherein said graphically displaying uses a pie graph to graphically display the execution profile information of each routine.

9. A method according to claim 1, wherein said graphically displaying uses a radar chart to graphically display the execution profile information of each routine.

10. A method according to claim 1, further comprising:

accepting a designation of an item of the execution profile information displayed by said graphically displaying, and

displaying a part of source code of the parallel processing program which corresponds to the item designated in the execution profile information.

11. A computer-readable medium encoded with a computer program in which the method as recited in claim 1 is implemented.